Development and adoption of MCP

From: aidotengineer

The Model Context Protocol (MCP) has rapidly gained traction in the AI engineering world, with a dedicated track at the 2025 AI Engineer World’s Fair [00:21:49]. Described as the “default protocol for connecting with the outside world” [00:29:09], MCP is seen as a foundational shift in the internet’s economy, where “tool calls are becoming the new clicks” [02:31:08].

Origin Story and Evolution

The concept of MCP originated in mid-2023 with Anthropic co-creators David and Justin [02:33:05]. They observed engineers constantly copying and pasting context (like Slack messages or error logs) into an LLM’s context window [02:33:20]. Their vision was for LLMs, such as Claude, to “climb out of its box, reach out into the real world and bring that context and those actions to the model” [02:33:52].

The genesis of MCP was rooted in the idea of “model agency”—giving the model the ability to interact with the outside world [02:34:06]. They concluded that an open-source, standardized protocol was essential for this to scale [02:34:20]. This approach aimed to overcome the challenges of closed-source ecosystems, where integrating an LLM would require complex business development and interface alignment [02:34:29].

The protocol was initially developed by a small internal team at Anthropic and launched during their company hack week in November 2023 [02:35:28]. It quickly went viral internally, with engineers building MCPs to automate workflows [02:35:41]. This success led to MCP being open-sourced in November 2023 [02:36:08].

Despite the initial enthusiasm, the public launch faced questions about the necessity of a new, open-source protocol given existing model tool-calling capabilities [02:36:40]. However, a significant turning point came with Cursor’s adoption of MCP, followed by other coding tools like VS Code and Sourcegraph [02:37:18]. More recently, major AI players including Google, Microsoft, and OpenAI have also adopted MCP, solidifying its status as a standard [02:37:47].

Key Features and Principles

MCP’s core focus is on enabling model agency, allowing the model’s intelligence to choose actions and decide what to do [02:39:18]. The protocol is built on principles that anticipate a future with many more servers than clients [02:40:26], optimizing for server simplicity and better tooling for server builders [02:40:41]. This design choice means that complexity is often pushed to the client side [02:40:50].

Recent advancements in MCP include:

• Streamable HTTP Support: This enhances bidirectionality, crucial for agents to communicate with each other [02:39:57].
• Improved OAuth 2.1: Initially a significant hurdle, the OAuth implementation was corrected through community feedback and contributions, allowing for enterprise-grade authorization [02:41:23].
• Registry API: This future development will make it easier for models to discover MCPs not explicitly provided to them, further enhancing model agency [02:43:08].
• Elicitation: Servers can now request more information from end-users, enabling dynamic interactions for tasks like flight booking, where user preferences (e.g., cheapest vs. fastest) can be clarified [02:42:30].
• Developer Experience: Ongoing efforts include updating SDKs and improving the inspector debugging tool [02:42:06].

MCP facilitates “rich stateful interactions” [03:07:07], with features like:

• Dynamic Discovery: Servers can dynamically reveal or hide tools based on context, as demonstrated by a game master agent revealing a “battle” tool only when a monster is present [03:09:43].
• Resources: Allows servers to return references to files or data rather than the full content, which LLMs or users can then act upon [03:10:24]. This enables contextual understanding of environments (e.g., Python settings, installed packages) without constant prompting [03:10:56].
• Sampling: Allows an MCP server to request LLM completions from the client, useful for summarizing resources or formatting web content into markdown [03:11:54].

Adoption and Integration with AI Applications

MCP’s adoption is driven by its ability to simplify common integration problems for AI applications [02:54:13]. Microsoft, a presenting sponsor of the World’s Fair, was an early adopter, integrating MCP with over 1500 tools [00:48:01] and supporting it across its platforms, including GitHub Copilot and Azure AI Foundry [00:49:35]. The Microsoft team highlighted how MCP and A2A (Agent-to-Agent protocol) are essential for building the “agentic web,” where agents interact across different clouds, companies, and devices [03:08:04].

Anthropic itself uses MCP internally at scale to manage tool calling and integrations [02:50:47]. They developed an “MCP gateway” to provide a single point of entry for engineers, standardizing authentication, rate limiting, and observability [02:58:07]. This “pit of success” model makes the “right thing to do the easiest thing to do,” ensuring consistent and secure integration [02:57:51].

Sentry also adopted MCP for its B2B SaaS business, focusing on integrating its application monitoring capabilities with agent architectures [03:21:06]. David Kramer, Sentry’s founder, emphasized that MCP is not merely an API wrapper; it requires careful design to provide context to agents effectively [03:26:04].

Challenges and Future Outlook

Despite rapid adoption, the MCP ecosystem faces problems and challenges.

Problems Faced by Developers

• “Copy and Paste Hell”: Early AI products suffered from a lack of connectivity, forcing users to manually copy and paste between applications [02:30:00]. MCP aims to solve this by allowing AI to connect to the rest of the world [02:30:13].
• Initial Confusion: At launch, many developers did not understand the need for a new open protocol when models could already call tools [02:36:40].
• “API Wrapper Syndrome”: Many early MCP implementations simply wrapped existing API endpoints one-to-one [02:45:41], leading to poor results because models struggle to reason about giant, undifferentiated JSON payloads [02:56:01]. A human-readable format like Markdown is often more effective [03:30:12].
• Client Support Gaps: Inconsistent MCP support across different clients (e.g., IDEs, cloud desktops) [03:26:38]. While VS Code Insiders offers full OAuth and MCP support, other platforms may lag [03:28:56].
• Security Concerns: The standard IO interface for MCP raises security concerns, including prompt injection risks [03:28:34]. Organizations are advised to trust only verified MCP tools [03:28:46]. The “lethal trifecta” arises when an AI system with access to private data is exposed to malicious instructions and has exfiltration mechanisms [01:42:36].
• Cost Management: AI tool calls can become expensive if not carefully managed, especially with large token payloads [03:31:52].
• Debugging and Logging: Developers have reported difficulties with debugging and logging MCP servers [03:13:17], though VS Code now offers a dev mode with console and debugger attachment [03:13:30].
• Lack of Streaming Responses for Tools: A current limitation is the absence of streaming responses for tools, which complicates agent-to-agent communication and forces workarounds like polling [03:33:01].

Opportunities and Future Direction

Despite the problems, the MCP ecosystem is still “really early” with massive opportunities [02:45:04].

• Higher Quality Servers: There’s a strong need for more, higher-quality MCP servers [02:45:37]. Building effective servers means designing for three users: the end-user, the client developer, and the model [02:46:01].
• Vertical Expansion: While initial MCPs were often for developer tools, there’s a desire for expansion into other verticals like sales, finance, legal, and education [02:46:31].
• Simplifying Server Building: Given the anticipated proliferation of servers, tools to simplify their creation (hosting, testing, deployment, evaluation) are crucial for both enterprises and indie developers [02:46:52].
• Automated MCP Generation: A moonshot vision suggests that as model intelligence grows, they will be able to write their own MCPs on the fly [02:47:44].
• AI Security Tooling: Investment in AI security, observability, and auditing tools is critical as MCP enables applications to interact with real-world data [02:48:25].
• Community Registry: Efforts are underway to create a community registry for easier discovery of MCP servers [03:15:32].
• Agent-Centric Development: The overarching advice is to focus on building agents, treating MCP as a pluggable architecture for these services [03:32:41]. This approach emphasizes designing for context within specific workflows, allowing for greater control over model interactions and results [03:32:50].

The vision for MCP is to enable a transformative potential for AI applications by ensuring the ecosystem catches up to the capabilities already outlined in the protocol’s specification [03:18:11].